Primitive of the day · 2026-10-30

generalization-second-clean-round

Convergence needs two consecutive clean rounds where the SECOND is a generalization test on a fresh un-tuned slice -- a clean round on tuned material proves nothing.

When this applies

An iterative improve-and-verify loop is about to be declared converged.

Preconditions

• A fresh slice exists that was never used to tune the artifact

Do this

After the first clean round, run a second round on a fresh un-tuned slice/task; only two consecutive cleans with the second being the generalization test count as convergence.

What you should see

A convergence claim backed by a clean generalization round, not just a clean tuned round.

How it fails if ignored

Both clean rounds run on tuned material; goalpost-moving after a failed generalization round.

Do not use when

Loops with no held-out material available at all (then state that limitation explicitly instead).

Kind: calibration. Part of the skill Bottomup topdown elicitation. Free to reuse in your own agent skills.